Error in unserialize(socklist[[n]]) : error reading from connection about kernelshap HOT 4 CLOSED

Hello and sorry you run into a problem.

I probably cannot solve the error coming from parallel computing on Windows. But maybe I can help you with the single-threaded problem (no movement in error bar):

1. SHAP analyses are usually done by decomposing 200-2000 predictions. Maybe you can start with an X consisting of 500 randomly sampled rows.
2. What is the size of your background data set? Scott Lundberg proposes to use small sets of only a few rows (up to 100). Also here I'd suggest to use a bg_X of 20. If the speed is acceptable, you can increase to 100.
3. I guess you are not using exact = TRUE?
4. Are you using the current CRAN version?

What do you observe?

from kernelshap.

Thank you for the swift answer. I was able to solve the problem thanks to your ideas and some tinkering. Initially, my bg_x and x were the same, both being 27k x 140. I set x to be a subset of 2000 rows and bg_x to be 500 with parallel processing for 12 threads enabled with exact = FALSE. Processing finished after about 5 minutes.

Will the fact that I am now using a smaller dataset for x and bg_x change my results relative to using my full dataset for x and bg_x? I am able to use the full dataset for x and bg_x with treeSHAP.

from kernelshap.

Sweet! Thanks for testing. With 500 rows, your background data is still very large, but 5 minutes is quite acceptable. Using even larger X and bg_X will change the result, but only slightly. I usually decompose between 1000 and 2000 predictions, even with TreeSHAP.

TreeSHAP is indeed magnitudes faster than KernelSHAP but can only be used for tree-based methods, while KernelSHAP works for all model classes. For trees, I would not use KernelSHAP in practice.

By default, in your case, exact = FALSE, so you don't need to specify it explicitly.

from kernelshap.

Since I'm doing research, a longer processing time is acceptable but I suppose for production TreeSHAP would be preferred. Thanks again!

from kernelshap.